Save 4 hours! We auto-setup your workflow! Free!

Drop your workflow.json — we handle every dependency, custom node, and model. Just open the link and run.

Auto-Setup Workflow Json (Free) Now!
ComfyUI > Nodes > ComfyUI_LC_Vision_nodes

ComfyUI Extension: ComfyUI_LC_Vision_nodes

Repo Name

ComfyUI_LC_Vision_nodes

Author
lonecatone23 (Account age: 0 days)
Nodes
View all nodes(1)
Latest Updated
2026-11-14
Github Stars
0.02K

How to Install ComfyUI_LC_Vision_nodes

Install this extension via the ComfyUI Manager by searching for ComfyUI_LC_Vision_nodes
  • 1. Click the Manager button in the main menu
  • 2. Select Custom Nodes Manager button
  • 3. Enter ComfyUI_LC_Vision_nodes in the search bar
After installation, click the Restart button to restart ComfyUI. Then, manually refresh your browser to clear the cache and access the updated list of nodes.

Visit ComfyUI Online for ready-to-use ComfyUI environment

  • Free trial available
  • 16GB VRAM to 80GB VRAM GPU machines
  • 400+ preloaded models/nodes
  • Freedom to upload custom models/nodes
  • 200+ ready-to-run workflows
  • 100% private workspace with up to 200GB storage
  • Dedicated Support

Run ComfyUI Online

ComfyUI_LC_Vision_nodes Description

ComfyUI_LC_Vision_nodes uses local vision models to caption images and video, refine prompts, review renders, plan video shots, and help diagnose ComfyUI workflows through an assistant sidebar.

ComfyUI_LC_Vision_nodes Introduction

ComfyUI_LC_Vision_nodes adds local vision and writing tools to ComfyUI. Its nodes can look at reference images or video, describe what they see, improve prompts, check finished renders, and plan shots for video workflows. A Vision Assistant sidebar also helps you understand and troubleshoot the workflow open on your canvas.

The extension is designed for workflows where you use a vision model repeatedly. The LC Vision Loader can keep a model ready between runs, while its memory controls let you make room for an image-generation model when needed. Caption and Moviemaker can consider several references in one request, with each reference identified separately.

The author keeps this extension separate from the companion ComfyUI_LC123_nodes pack. Some features below work on their own; others explicitly depend on LC123 or LC AV. The original README describes version 1.6.31, with 12 nodes and the Vision Assistant sidebar. It does not provide a version-by-version changelog.

How ComfyUI_LC_Vision_nodes Works

Think of the vision model as a local creative assistant. You give it instructions and, when useful, pictures or video frames. It returns text: a caption, a revised prompt, answers to quality-check questions, or a plan for video shots. It does not generate the final image or video by itself. You connect its text outputs to the other nodes in your ComfyUI workflow.

A typical image workflow looks like this:

  1. Load a model with LC Vision Loader.
  2. Show it a reference with LC Vision Caption, or give LC Vision Prompt Enhancer an idea to rewrite.
  3. Use the resulting text as a prompt for your image-generation workflow.
  4. Optionally check the render with LC Vision Reviewer before saving or further processing it.

The Loader pairs a language-model file with its matching mmproj file, which provides the model’s image-understanding capability. It lists local model pairs only when both files are present. It can also offer selected models for download from its dropdown or from LC Settings → Vision → Vision models.

The model needs working memory while it reads your references and writes a response. The Loader’s n_ctx setting controls how much context it can hold; longer or more complex requests need more context and generally reserve more VRAM. keep_model_loaded controls whether the model remains in memory after answering. Keeping it loaded saves time on later requests; unloading it leaves more room for image or video generation.

Some backend errors can occur when a model is reused. The Caption node first resets its context and, for a specific decode failure, reloads the model and retries once. This recovery is automatic, but it does not hide unrelated errors.

ComfyUI_LC_Vision_nodes Features

LC Vision Loader

The Loader supplies a model handle to the other LC Vision nodes.

  • Model selection: Choose an available local model pair or a supported Download: entry. After a download, the selection changes to the local filename for subsequent runs.
  • device: Choose auto, cuda, or cpu. A CPU first test can help you establish that the workflow runs without competing for GPU memory, though it will be slower.
  • n_ctx: The available choices are 4096, 8192, 16384, and 32768; the default is 8192. As a starting point, use 4096 for one image and a short request, 8192 for Caption or Prompt Enhancer, and a larger value for several references or video frames. Higher settings reserve substantially more VRAM.
  • keep_model_loaded: On by default for faster repeated use. Turn it off when your image or video model needs the memory; LC Vision will load again for its next request.
  • n_gpu_layers: -1 offloads all possible layers to the GPU; 0 runs on the CPU; another number offloads that many layers.
  • n_batch: Controls processing batch size. The node also adjusts the related internal batch setting, n_ubatch, to match.

With LC123 installed and its System & Model Optimization Report run, the Loader may show a ★ model suggestion suited to your graphics card. It is advice, not an automatic change to your saved selection.

LC Vision Caption

Caption turns a prompt and visual references into descriptive text. It accepts up to three separate reference images and one separate reference video, labeling them distinctly so the model can discuss the right source.

Use max_image_side to reduce oversized references before they reach the model. This can help prevent capacity errors. Use max_tokens to request a shorter or longer answer; the node aims to finish on a complete sentence rather than leave a caption abruptly cut off. A shared style_tag setting can steer the wording toward styles such as Realistic, Anime, Cinematic, Fantasy, Noir, or Illustration, even when the reference itself has a different look. Leave it at None when you want the description to follow the reference more closely.

LC Vision Reviewer (Beta)

Reviewer is a quality checkpoint for a finished image. It asks the vision model yes-or-no questions about hands, faces, readable text, and visual artifacts. If you connect your positive prompt, it can also check whether the result appears to match the request.

Add your own questions in custom_checks, one per line, phrased so “yes” means the result is good. For example: “Is she wearing a parka?” Its outputs include the image, a passed result, and a written report.

With on_fail set to flag (the default), the image continues through the workflow with a reported issue. Set it to block to stop connected downstream steps, such as saving or upscaling, when a check fails. For batches larger than four frames, Reviewer examines selected frames together as a video. Treat its findings as a shortlist for your own review: a vision model can miss mistakes or flag something that is acceptable.

LC Vision Prompt Enhancer

Prompt Enhancer rewrites text using the loaded model. You can give it a written idea alone or attach optional image_1 and image_2 references.

Its presets include Enhance, Refine, Creative Rewrite, Detailed Visual, Character Anchor, and four Flux 3 Klein options: Flux 3 Text to Image, Flux 3 Edit, Flux 3 Layout (boxes), and Flux 3 Edit (boxes). Choose Custom to supply your own system prompt through custom_system_prompt; an empty custom prompt falls back to Enhance. The edit presets expect the picture being edited in image_1.

Character Anchor is useful when you want a consistent person across multiple images. From references and a few notes, it describes enduring traits such as face, hair, eyes, skin, and build, while leaving clothing, pose, background, and lighting open for each scene.

When used with LC MOSS Layout Prompt, Prompt Enhancer can preserve a scene-and-box layout while improving the scene and individual box descriptions. Related LC123 and LC AV integrations can use the vision model to propose bounding boxes from an image or idea, or describe a room picture for a 360-degree scene. These integrations depend on the named companion packs.

LC Vision Danbooru Caption

Danbooru Caption creates short prompts aimed at SDXL, Pony, and Illustrious workflows. In Image analysis mode, it works from an image; in Prompt enhance mode, it rewrites the idea in the prompt box.

Choose an output format to suit your model: short caption, detailed caption, Danbooru tags, or prompt gen mixed (brief prose followed by tags). Tag output is checked against a real tag list; unsupported tags are repaired or removed where possible. max_tokens controls approximate length, and style_tag can add appropriate style terms.

Optional quality_tags presets are Illustrious, NoobAI, and Pony; the default is none. Try them only when they suit the model you use, since extra quality terms are not always helpful. The node provides separate positive and negative outputs. Your own negative text comes first, followed by its automatic negative terms with duplicates removed. Its single image slot and relatively short outputs are intentional; use Caption or Prompt Enhancer for longer, more elaborate descriptions.

LC Vision Moviemaker

Moviemaker writes a plan for multiple independently generated video segments in one model request. Because it sees the whole story at once, it can describe how one segment continues into the next while returning a separate prompt for each.

Set your story, length_seconds, number of segments, and a preset. Presets cover three MiniMax H3 modes, T2V, R2VA, and FL2VA, each with an NSFW variant; Custom lets you change the creative instructions while retaining the node’s structural planning rules. The output sockets grow with the segment count.

You can provide four reference images and one reference video. Images are associated with labeled pictures and shots; the video informs motion or style continuity. Each segment is planned on its own local clock, and the node assembles the overall timeline afterward. end_pad_seconds defaults to 0.5; increasing it can help when generated plans consistently end their segments too early. A [Shot N] label identifies a reference, not an instruction to cut the camera.

LC Vision Showrunner, Shot List, and Storyboard (Beta)

These tools support longer video projects:

  • Showrunner plans a chain of connected segments. It repeats one style description across them, plans overlapping handoffs, and writes a plan for LC H3 Make in LC AV to render and join. It saves project settings and plans, allowing unchanged plans to be reused without another model run. extend_segments can give a shot with a long spoken line more time.
  • Shot List lets you enter a place, cast, and one row per shot rather than write a full story from scratch. It produces a story for Showrunner, suggests a segment count, and warns about issues such as an unknown speaker or an overly long line. No AI model is involved in filling out the Shot List.
  • Storyboard Builder identifies missing face, place, and opening-scene reference pictures for a project and prepares prompts to make them in a separate image workflow. Storyboard Save adds the resulting pictures to the project. Storyboard Sheet lets you review, reuse, regenerate, or revise a picture’s prompt from its card. Storyboard pictures are optional; a Showrunner movie can run without them.

LC Vision Redirect (Beta) remains available so older workflows can open. Showrunner now handles the join correction that Redirect previously provided.

LC Note Translation

When LC123 is available, LC Vision can translate an LC Note and its title through Translate now. This is an integration, not a node you need to wire. It loads a local model for the translation and unloads it afterward. Links, code, filenames, and numbers are protected and checked, but important translations should still be proofread. Your manual translation edits are retained.

Vision Assistant Sidebar

Open the LC logo tab to use the assistant with the workflow currently on your canvas, including workflows that do not use LC Vision nodes.

  • Checkup runs without loading a model. It can point out missing packs or models, questionable settings, damaged connections, and recent errors in plain language. Some issues offer a direct fix. With the relevant LC companion pack installed, it can also offer conversions to equivalent LC nodes or LC Notes.
  • Chat answers questions such as “What’s broken?” or “Where do I put my image?” Selecting a node first gives it a specific subject to explain. It asks for confirmation before making a change and provides Undo.
  • Guided tour walks through a workflow on the canvas, highlighting relevant nodes and settings. Workflow authors can save a tour into the workflow file.
  • Movie appears for supported LC AV movie workflows and shows segment previews and checks, with options to redo a segment or continue rendering.
  • Batch snapshot warns before a run that appears unusually large or accidental, such as many images or repeated fixed-seed runs. An ordinary single-image run proceeds without interruption.
  • Snapshots and Library save and restore selected workflow settings. Prompts, seeds, and input images are excluded unless you choose to include them.
  • Models in this workflow lists files used by the workflow, whether they are available, and which nodes use them. Download-link and CivitAI lookup controls are available when you choose to use them.

The Assistant’s Auto chat setting selects a model based on available hardware. Its model unloads after inactivity or when you start a run. In LC Settings → Vision, you can adjust the assistant, chat model, batch snapshot, default context for new Loaders, and hidden tips or swap suggestions. The panel’s ? button opens its fuller help.

ComfyUI_LC_Vision_nodes Models

The Loader works with supported Qwen vision-language GGUF models and their matching mmproj vision files. The original README identifies these curated choices:

Model choiceWhen to consider it
Qwen3-VL 4BA smaller choice that leaves more room for multi-frame requests, Moviemaker, or other models sharing the GPU.
Qwen3-VL 8BThe author’s preferred choice for captioning and image-detail fidelity when you have enough memory.
Qwen3.5-VL 4B Q8An alternative in the curated download list.
Qwen3.5-VL 9B Q4_K_M or Q8Larger alternatives in the curated download list. The author reports that Qwen3.5 was slower and guessed more in the author’s tests, so Qwen3-VL remains the suggested starting point.

The curated Qwen3-VL 4B and 8B downloads include abliterated variants and Q8_0 or f16 options. These labels describe different stored versions of a model, not different node features. As a broad starting point, Q8_0 is intended for lower-VRAM or new setups, while f16 is aimed at well-equipped machines or cloud rigs. Available local models can also depend on the GGUF/mmproj pairs you have placed in folders recognized by ComfyUI.

There is no single best model for every workflow. Start with a smaller choice when memory is tight or you are processing several frames; try a larger choice when image detail matters and you have room. The Loader’s ★ suggestion, when available through LC123’s report, can provide a machine-specific hint without changing your selection.

Troubleshooting ComfyUI_LC_Vision_nodes

The LC Vision nodes or sidebar do not appear

  1. Restart ComfyUI after adding the extension; new nodes need a restart to be discovered.
  2. Confirm that the extension folder is not nested inside an extra folder. In the original README, __init__.py is expected directly inside ComfyUI/custom_nodes/ComfyUI_LC_Vision_nodes/.
  3. After an update to the sidebar or other web interface, hard-refresh your browser.
  4. Check the ComfyUI console for the extension’s load message or an error.

The Loader says llama_cpp is missing or cannot load

The extension requires a vision-capable build from the JamePeng llama-cpp-python fork, rather than the ordinary PyPI build. The error message includes the Python path ComfyUI is using and, in brackets, the underlying import error. Keep the full error line when seeking help: it identifies both the environment and the likely cause. The repository’s original documentation has detailed environment-specific troubleshooting.

If the message says the installed copy is the original build, another extension may have replaced the vision-capable fork. If it names a missing small dependency such as jinja2, follow the specific dependency guidance printed in the error and restart ComfyUI. Prebuilt fork wheels described in the original README target NVIDIA CUDA and Apple Silicon Metal; AMD, Intel, and CPU-only environments have additional backend requirements.

A model is missing from the Loader’s list

  1. Check that the model .gguf and its matching mmproj .gguf are both present in the same folder recognized by the Loader.
  2. Look for a supported Download: choice in the Loader or open LC Settings → Vision → Vision models.
  3. If you opened someone else’s workflow, use the Vision Assistant’s Models in this workflow or Checkup to see whether the file is missing, moved, or known under another name.

A request runs out of memory or reports a capacity problem

  1. For image references, lower Caption’s max_image_side.
  2. Use a smaller Loader n_ctx when the request contains only one image and a short prompt.
  3. Turn off keep_model_loaded if an image or video model needs the same GPU memory afterward.
  4. Consider a smaller model for multi-frame work.

A caption or prompt is too short, too long, or follows the wrong style

Check max_tokens first: a low value deliberately produces a brief answer. Then check style_tag. A selected style can steer the description away from the reference’s actual look; choose None when accuracy to the image is more important. For Danbooru Caption, also check that mode and output match what you intend: image analysis ignores the idea in the prompt box, while prompt enhance works from that idea.

The console says a reused model was rebuilt after a decode failure

A message that LC Vision is rebuilding and retrying once means its automatic recovery has caught the specific backend decode issue described by the author. You do not need to change your prompt for a successful retry. If the retry also fails, keep the full console error when reporting the problem.

Frequently asked questions

Does LC Vision create images or render movies?
No. Its core vision nodes produce text or review results for other nodes to use. The Showrunner plan is rendered and joined by the appropriate LC AV workflow.

Do I need the companion packs?
Not for every feature. The core Loader, Caption, Prompt Enhancer, Reviewer, Danbooru Caption, Moviemaker, and general Vision Assistant functions are described as part of this extension. Specific LC Note, bounding-box, report, and H3 movie integrations require the named companion packs.

Can Reviewer guarantee a flawless image?
No. It can help you spot likely issues, but you should inspect important outputs yourself.

Why does the model take time to start again?
If keep_model_loaded is off, it unloads after answering to free memory. The next LC Vision request must load it again.

What should I try for my first workflow?
Use the example workflow below, select an available model pair, and start with a short caption request and one reference image. The author suggests testing the Loader with device: cpu first, then moving to cuda or auto once the basic workflow runs.

Learn More about ComfyUI_LC_Vision_nodes

  • Extension repository: The primary place to find the extension’s documentation, report a reproducible problem, and inspect its git tags for release history. The original README describes the current pack rather than listing individual version changes.
  • Example workflows: The repository’s workflows/ folder includes LC Vision LLM Modules, with Caption, Prompt Enhancer, Moviemaker, and Reviewer wired together. Replace its example model and references with your own.
  • Vision Assistant help: Click ? in the sidebar panel for its full in-app guidance.
  • ComfyUI: Background on the node-based application in which these workflows run.
  • ComfyUI_LC123_nodes and ComfyUI_LC_AV_nodes: Companion packs for the LC integrations and movie-rendering workflows described above.

ComfyUI_LC_Vision_nodes Related Nodes

RunComfy
Copyright 2025 RunComfy. All Rights Reserved.

RunComfy is the premier ComfyUI platform, offering ComfyUI online environment and services, along with ComfyUI workflows featuring stunning visuals. RunComfy also provides AI Models, enabling artists to harness the latest AI tools to create incredible art.